Bitcoin’s notoriously volatile price may look like it bounces around arbitrarily but an MIT computer scientist has developed a machine-learning algorithm that appears to have found a pattern in the mayhem, helping him to nearly double his investment in less than two months.
MIT associate professor Devavrat Shah and his student Kang Zhang set out to see if the Bayesian regression statistical technique could be used to enable an algorithm to spot patterns in bitcoin trading data which could then help predict future prices.
The team collected five months’ worth of price data – more than 200 million data points – from the popular Okcoin bitcoin exchange in China. Once the data was fed into the model they used it to predict average price movement over 10 second periods and used this information to decide whether to buy, sell or do nothing.
During a 50 day period, the team made 2872 simulation trades and scored a return on their investment of 89% with a Sharpe ratio of 4.1, indicating a strong return in relation to the risk taken.
Bitcoin’s price is known for fluctuating wildly. At the time of publication, the cryptocurrency is trading at around $360 having taken a sharp dive in recent weeks. However, Shah, who previously used the Bayesian regression technique to predict Twitter trending topics, is confident that any quantity that changes over time can be modelled.
He tells MIT News: “We developed this method of latent-source modelling, which hinges on the notion that things only happen in a few different ways. Instead of making subjective assumptions about the shape of patterns, we simply take the historical data and plug it into our predictive model to see what emerges.”